Automatic assessment of posture deviations in assembly tasks

Abstract

The aim of this study is to investigate the development and the evaluation of a computer vision-based framework to aid the automatic assessment of posture deviations in assembly tasks in realistic work environments. A posture deviation refers to a time-varying working posture performed by the worker, that deviates from ergonomically safe body postures expected in the context of particular work tasks and is known to impose increased physical strain. The estimation of their occurrences can serve as indicators, known as risk factors, for the assessment of physical ergonomics towards the prevention of physical strain and in the-long-term of work-related musculo-skeletal disorders (WMSD). Using visual information acquired by camera sensors, our goal is to estimate the full body motion of a line worker in 3D space, unobtrusively, and to perform classification of four types of posture deviations, also noted as ergonomically sub-optimal working postures that were employed by the MURI risk analysis tool. We formulate a learning-based action classification task using Deep Graph-based Neural Networks and differential temporal alignment cost as a classification measure to estimate the type and risk level of the observed posture deviation during work activities. To evaluate the efficiency of the proposed approach, a new video dataset was captured in the context of the sustAGE project, that demonstrate two different workers during car door assembly actions in a simulated production line in an actual workplace. Rich annotation data were provided by experts in manufacturing and ergonomics. Both quantitative and qualitative evaluation of the proposed framework provide evidence for its efficiency and reliability in supporting ergonomic risk assessment and preventive actions for WMSD in real working environments.